oota-llvm.git
11 years agoRemove unused function.
Rafael Espindola [Mon, 15 Apr 2013 15:13:10 +0000 (15:13 +0000)]
Remove unused function.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179530 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMake the host endianness check an integer constant expression.
Rafael Espindola [Mon, 15 Apr 2013 14:44:24 +0000 (14:44 +0000)]
Make the host endianness check an integer constant expression.

I will remove the isBigEndianHost function once I update clang.

The ifdef logic is designed to
* not use configure/cmake to avoid breaking -arch i686 -arch ppc.
* default to little endian
* be as small as possible

It looks like sys/endian.h is the preferred header on most modern BSD systems,
but it is better to change this in a followup patch as machine/endian.h is
available on FreeBSD, OpenBSD, NetBSD and OS X.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179527 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoReplace uses of the deprecated std::auto_ptr with OwningPtr.
Andy Gibbs [Mon, 15 Apr 2013 12:06:32 +0000 (12:06 +0000)]
Replace uses of the deprecated std::auto_ptr with OwningPtr.

This is a rework of the broken parts in r179373 which were subsequently reverted in r179374 due to incompatibility with C++98 compilers.  This version should be ok under C++98.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179520 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoEnable all targets by default on Visual Studio.
Tim Northover [Mon, 15 Apr 2013 11:53:05 +0000 (11:53 +0000)]
Enable all targets by default on Visual Studio.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179518 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert "Recommit r179497 after fixing uninitialized variable." until
Eric Christopher [Mon, 15 Apr 2013 07:31:37 +0000 (07:31 +0000)]
Revert "Recommit r179497 after fixing uninitialized variable." until
I can fix the testcases here:

http://lab.llvm.org:8011/builders/clang-native-arm-cortex-a9/builds/6952

This reverts commit r179512 due to testcases specifying triples
that they didn't actually mean and causing failures on other platforms.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179513 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRecommit r179497 after fixing uninitialized variable.
Eric Christopher [Mon, 15 Apr 2013 07:07:21 +0000 (07:07 +0000)]
Recommit r179497 after fixing uninitialized variable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179512 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDocument our desire to enable the loop vectorizer on -Os in future releases.
Nadav Rotem [Mon, 15 Apr 2013 05:56:55 +0000 (05:56 +0000)]
Document our desire to enable the loop vectorizer on -Os in future releases.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179511 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDocs: merge the description of the BB and SLP vectorizers and document the -fslp...
Nadav Rotem [Mon, 15 Apr 2013 05:53:23 +0000 (05:53 +0000)]
Docs: merge the description of the BB and SLP vectorizers and document the -fslp-vectorize-aggressive flag.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179510 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd an option -vectorize-slp-aggressive for running the BB vectorizer. Make -fslp...
Nadav Rotem [Mon, 15 Apr 2013 05:39:58 +0000 (05:39 +0000)]
Add an option -vectorize-slp-aggressive for running the BB vectorizer. Make -fslp-vectorize run the slp-vectorizer.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179508 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRename the slp-vectorizer clang/llvm flags. No functionality change.
Nadav Rotem [Mon, 15 Apr 2013 04:54:42 +0000 (04:54 +0000)]
Rename the slp-vectorizer clang/llvm flags. No functionality change.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179505 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSLPVectorizer: Add support for vectorizing trees that start at compare instructions.
Nadav Rotem [Mon, 15 Apr 2013 04:25:27 +0000 (04:25 +0000)]
SLPVectorizer: Add support for vectorizing trees that start at compare instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179504 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agofix include path in doc Extending LLVM
Jia Liu [Mon, 15 Apr 2013 03:26:13 +0000 (03:26 +0000)]
fix include path in doc Extending LLVM

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179503 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMark all PPC comparison instructions as not having side effects
Hal Finkel [Mon, 15 Apr 2013 02:37:46 +0000 (02:37 +0000)]
Mark all PPC comparison instructions as not having side effects

Now that the CR spilling issues have been resolved, we can remove the
unmodeled-side-effect attributes from the comparison instructions (and also
mark them as isCompare). By allowing these, by default, to have unmodeled side
effects, we were hiding problems with CR spilling; but everything seems much
happier now.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179502 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix PPC64 CR spill location for callee-saved registers
Hal Finkel [Mon, 15 Apr 2013 02:07:05 +0000 (02:07 +0000)]
Fix PPC64 CR spill location for callee-saved registers

This fixes an ABI bug for non-Darwin PPC64. For the callee-saved condition
registers, the spill location is specified relative to the stack pointer (SP +
8). However, this is not relative to the SP after the new stack frame is
established, but instead relative to the caller's stack pointer (it is stored
into the linkage area of the parent's stack frame).

So, like with the link register, we don't directly spill the CRs with other
callee-saved registers, but just mark them to be spilled during prologue
generation.

In practice, this reverts r179457 for PPC64 (but leaves it in place for PPC32).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179500 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert "Remove some unused triple and data layout."
Eric Christopher [Sun, 14 Apr 2013 23:35:36 +0000 (23:35 +0000)]
Revert "Remove some unused triple and data layout."

This reverts commit r179497 and the accompanying commit as it broke random platforms that aren't osx.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179499 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove some unused triple and data layout.
Eric Christopher [Sun, 14 Apr 2013 23:32:44 +0000 (23:32 +0000)]
Remove some unused triple and data layout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179498 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoIf we've specified a triple on the command line then go ahead
Eric Christopher [Sun, 14 Apr 2013 23:32:40 +0000 (23:32 +0000)]
If we've specified a triple on the command line then go ahead
and use that as the default triple for the module and target
data layout.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179497 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse object file specific section type for initial text section
Nico Rieck [Sun, 14 Apr 2013 21:18:36 +0000 (21:18 +0000)]
Use object file specific section type for initial text section

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179494 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoReorders two transforms that collide with each other
David Majnemer [Sun, 14 Apr 2013 21:15:43 +0000 (21:15 +0000)]
Reorders two transforms that collide with each other

One performs: (X == 13 | X == 14) -> X-13 <u 2
The other: (A == C1 || A == C2) -> (A & ~(C1 ^ C2)) == C1

The problem is that there are certain values of C1 and C2 that
trigger both transforms but the first one blocks out the second,
this generates suboptimal code.

Reordering the transforms should be better in every case and
allows us to do interesting stuff like turn:
  %shr = lshr i32 %X, 4
  %and = and i32 %shr, 15
  %add = add i32 %and, -14
  %tobool = icmp ne i32 %add, 0

into:
  %and = and i32 %X, 240
  %tobool = icmp ne i32 %and, 224

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179493 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMake the command line triple match the module triple.
Nadav Rotem [Sun, 14 Apr 2013 20:13:05 +0000 (20:13 +0000)]
Make the command line triple match the module triple.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179492 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMiscellaneous cleanups for VecUtils.h
Benjamin Kramer [Sun, 14 Apr 2013 09:33:08 +0000 (09:33 +0000)]
Miscellaneous cleanups for VecUtils.h

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179483 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDocument the SLP infrastructure.
Nadav Rotem [Sun, 14 Apr 2013 07:42:25 +0000 (07:42 +0000)]
Document the SLP infrastructure.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179480 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSLP: Document the scalarization cost method.
Nadav Rotem [Sun, 14 Apr 2013 07:22:22 +0000 (07:22 +0000)]
SLP: Document the scalarization cost method.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179479 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDocument the decision to assume that the cost of floats is twice as much as integers.
Nadav Rotem [Sun, 14 Apr 2013 05:55:18 +0000 (05:55 +0000)]
Document the decision to assume that the cost of floats is twice as much as integers.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179478 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse i32 for all SPARC shift amounts, even in 64-bit mode.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 05:48:50 +0000 (05:48 +0000)]
Use i32 for all SPARC shift amounts, even in 64-bit mode.

Test case by llvm-stress.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179477 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove unused function attributes.
Nadav Rotem [Sun, 14 Apr 2013 05:47:04 +0000 (05:47 +0000)]
Remove unused function attributes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179476 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSLPVectorizer: Add support for trees that don't start at binary operators, and add...
Nadav Rotem [Sun, 14 Apr 2013 05:15:53 +0000 (05:15 +0000)]
SLPVectorizer: Add support for trees that don't start at binary operators, and add the cost of extracting values from the roots of the tree.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179475 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for the abs64 SPARC v9 code model.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 05:10:36 +0000 (05:10 +0000)]
Add support for the abs64 SPARC v9 code model.

For when 16 TB just isn't enough.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179474 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for the SPARC v9 abs44 code model.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 04:57:51 +0000 (04:57 +0000)]
Add support for the SPARC v9 abs44 code model.

This is the default model for non-PIC 64-bit code. It supports
text+data+bss linked anywhere in the low 16 TB of the address space.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179473 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse target flags for printing SPARC asm operands.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 04:35:19 +0000 (04:35 +0000)]
Use target flags for printing SPARC asm operands.

64-bit code models need multiple relocations that can't be inferred from
the opcode like they can in 32-bit code.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179472 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAlso put target flags on SPARC constant pool references.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 04:35:16 +0000 (04:35 +0000)]
Also put target flags on SPARC constant pool references.

Constant pool entries are accessed exactly the same way as global
variables.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179471 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSLPVectorizer: add initial support for reduction variable vectorization.
Nadav Rotem [Sun, 14 Apr 2013 03:22:20 +0000 (03:22 +0000)]
SLPVectorizer: add initial support for reduction variable vectorization.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179470 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix patterns for 64-bit pointers.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 01:53:23 +0000 (01:53 +0000)]
Fix patterns for 64-bit pointers.

This fixes the pic32 code model for SPARC v9.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179469 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd target flags to SPARC address operands.
Jakob Stoklund Olesen [Sun, 14 Apr 2013 01:33:32 +0000 (01:33 +0000)]
Add target flags to SPARC address operands.

SDNodes and MachineOperands get target flags representing the %hi() and
%lo() assembly annotations that eventually become relocations.

Also define flags to be used by the 64-bit code models.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179468 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMark all PPC CR registers to be spilled as live-in and tag MFCR appropriately
Hal Finkel [Sat, 13 Apr 2013 23:06:15 +0000 (23:06 +0000)]
Mark all PPC CR registers to be spilled as live-in and tag MFCR appropriately

Leaving MFCR has having unmodeled side effects is not enough to prevent
unwanted instruction reordering post-RA. We could probably apply a stronger
barrier attribute, but there is a better way: Add all (not just the first) CR
to be spilled as live-in to the entry block, and add all CRs to the MFCR
instruction as implicitly killed.

Unfortunately, I don't have a small test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179465 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDefine SPARC code models.
Jakob Stoklund Olesen [Sat, 13 Apr 2013 19:02:23 +0000 (19:02 +0000)]
Define SPARC code models.

Currently, only abs32 and pic32 are implemented. Add a test case for
abs32 with 64-bit code. 64-bit PIC code is currently broken.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179463 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoUse the correct types when matching ADDRri patterns from frame indexes.
Jakob Stoklund Olesen [Sat, 13 Apr 2013 19:02:16 +0000 (19:02 +0000)]
Use the correct types when matching ADDRri patterns from frame indexes.

It doesn't seem like anybody is checking types this late in isel, so no
test case.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179462 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoGlobalDCE: Fix an oversight in my last commit that could lead to crashes.
Benjamin Kramer [Sat, 13 Apr 2013 16:11:14 +0000 (16:11 +0000)]
GlobalDCE: Fix an oversight in my last commit that could lead to crashes.

There is a Constant with non-constant operands: blockaddress.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179460 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a scalability issue with complex ConstantExprs.
Benjamin Kramer [Sat, 13 Apr 2013 12:53:18 +0000 (12:53 +0000)]
Fix a scalability issue with complex ConstantExprs.

This is basically the same fix in three different places. We use a set to avoid
walking the whole tree of a big ConstantExprs multiple times.

For example: (select cmp, (add big_expr 1), (add big_expr 2))
We don't want to visit big_expr twice here, it may consist of thousands of
nodes.

The testcase exercises this by creating an insanely large ConstantExprs out of
a loop. It's questionable if the optimizer should ever create those, but this
can be triggered with real C code. Fixes PR15714.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179458 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSpill and restore PPC CR registers using the FP when we have one
Hal Finkel [Sat, 13 Apr 2013 08:09:20 +0000 (08:09 +0000)]
Spill and restore PPC CR registers using the FP when we have one

For functions that need to spill CRs, and have dynamic stack allocations, the
value of the SP during the restore is not what it was during the save, and so
we need to use the FP in these cases (as for all of the other spills and
restores, but the CR restore has a special code path because its reserved slot,
like the link register, is specified directly relative to the adjusted SP).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179457 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFurther generalize this scheduler test.
Andrew Trick [Sat, 13 Apr 2013 07:37:27 +0000 (07:37 +0000)]
Further generalize this scheduler test.

The order of copies depends on queue order, which is not very stable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179456 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a dislexic regex.
Andrew Trick [Sat, 13 Apr 2013 07:29:21 +0000 (07:29 +0000)]
Fix a dislexic regex.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179455 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd a missing REQUIRES: asserts
Andrew Trick [Sat, 13 Apr 2013 06:12:46 +0000 (06:12 +0000)]
Add a missing REQUIRES: asserts

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179453 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMI-Sched: DEBUG formatting.
Andrew Trick [Sat, 13 Apr 2013 06:07:49 +0000 (06:07 +0000)]
MI-Sched: DEBUG formatting.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179452 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMI-Sched cleanup. If an instruction has no valid sched class, do not attempt to check...
Andrew Trick [Sat, 13 Apr 2013 06:07:45 +0000 (06:07 +0000)]
MI-Sched cleanup. If an instruction has no valid sched class, do not attempt to check for a variant.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179451 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoX86 machine model: reduce SandyBridge and Haswell ILPWindow.
Andrew Trick [Sat, 13 Apr 2013 06:07:43 +0000 (06:07 +0000)]
X86 machine model: reduce SandyBridge and Haswell ILPWindow.

The initial values were arbitrary. I want them to be more
conservative. This represents the number of latency cycles hidden by
OOO execution. In practice, I think it should be within a small factor
of the complex floating point operation latency so the scheduler can
make some attempt to hide latency even for smallish blocks.

These are by no means the best values, just a starting point for
tuning heuristics. Some benchmarks such as TSVC run faster with this
lower value for SandyBridge. I haven't run anything on Haswell, but
it's shouldn't be 2x SB.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179450 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoMI-Sched: schedule physreg copies.
Andrew Trick [Sat, 13 Apr 2013 06:07:40 +0000 (06:07 +0000)]
MI-Sched: schedule physreg copies.

The register allocator expects minimal physreg live ranges. Schedule
physreg copies accordingly. This is slightly tricky when they occur in
the middle of the scheduling region. For now, this is handled by
rescheduling the copy when its associated instruction is
scheduled. Eventually we may instead bundle them, but only if we can
preserve the bundles as parallel copies during regalloc.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179449 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCatch another case where SD fails to propagate node order.
Andrew Trick [Sat, 13 Apr 2013 06:07:36 +0000 (06:07 +0000)]
Catch another case where SD fails to propagate node order.

I need to handle this for the test case in my following scheduler
commit.

Work is already under way to redesign the mechanism for node order
propagation because this case by case approach is unmaintainable.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179448 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd typenames to see if bot goes green.
Rafael Espindola [Sat, 13 Apr 2013 02:31:34 +0000 (02:31 +0000)]
Add typenames to see if bot goes green.

I hope this brings http://lab.llvm.org:8011/builders/clang-x86_64-darwin11-self-mingw32 back.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179446 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Move MipsTargetLowering::lowerINTRINSIC_W_CHAIN and
Akira Hatanaka [Sat, 13 Apr 2013 02:13:30 +0000 (02:13 +0000)]
[mips] Move MipsTargetLowering::lowerINTRINSIC_W_CHAIN and
lowerINTRINSIC_WO_CHAIN into MipsSETargetLowering.

No functionality changes.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179444 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSome versions of gcc don't like typenames in these places.
Rafael Espindola [Sat, 13 Apr 2013 01:55:34 +0000 (01:55 +0000)]
Some versions of gcc don't like typenames in these places.

Should fix the bots.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179441 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFinish templating MachObjectFile over endianness.
Rafael Espindola [Sat, 13 Apr 2013 01:45:40 +0000 (01:45 +0000)]
Finish templating MachObjectFile over endianness.

We are now able to handle big endian macho files in llvm-readobject. Thanks to
David Fang for providing the object files.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179440 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Reapply r179420 and r179421.
Akira Hatanaka [Sat, 13 Apr 2013 00:55:41 +0000 (00:55 +0000)]
[mips] Reapply r179420 and r179421.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179434 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Override TargetLoweringBase::isShuffleMaskLegal.
Akira Hatanaka [Sat, 13 Apr 2013 00:45:02 +0000 (00:45 +0000)]
[mips] Override TargetLoweringBase::isShuffleMaskLegal.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179433 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Simplify the logic by using parsePrimaryExpr. No functional
Chad Rosier [Fri, 12 Apr 2013 23:03:20 +0000 (23:03 +0000)]
[ms-inline asm] Simplify the logic by using parsePrimaryExpr.  No functional
change intended.  Test case previously added in r178568.
Part of rdar://13611297

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179425 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert r179420 and r179421.
Akira Hatanaka [Fri, 12 Apr 2013 22:40:07 +0000 (22:40 +0000)]
Revert r179420 and r179421.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179422 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] Instruction selection patterns for carry-setting and using add
Akira Hatanaka [Fri, 12 Apr 2013 22:24:52 +0000 (22:24 +0000)]
[mips] Instruction selection patterns for carry-setting and using add
instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179421 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[mips] v4i8 and v2i16 add, sub and mul instruction selection patterns.
Akira Hatanaka [Fri, 12 Apr 2013 22:14:24 +0000 (22:14 +0000)]
[mips] v4i8 and v2i16 add, sub and mul instruction selection patterns.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179420 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert r179409 because it caused some warnings and some of the build bots fail.
Nadav Rotem [Fri, 12 Apr 2013 22:02:26 +0000 (22:02 +0000)]
Revert r179409 because it caused some warnings and some of the build bots fail.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179418 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoInstCombine: Check the operand types before merging fcmp ord & fcmp ord.
Benjamin Kramer [Fri, 12 Apr 2013 21:56:23 +0000 (21:56 +0000)]
InstCombine: Check the operand types before merging fcmp ord & fcmp ord.

Fixes PR15737.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179417 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSLPVectorizer: add support for vectorization of diamond shaped trees. We now perform...
Nadav Rotem [Fri, 12 Apr 2013 21:16:54 +0000 (21:16 +0000)]
SLPVectorizer: add support for vectorization of diamond shaped trees. We now perform a preliminary traversal of the graph to collect values with multiple users and check where the users came from.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179414 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoCostModel: increase the default cost of supported floating point operations from...
Nadav Rotem [Fri, 12 Apr 2013 21:15:03 +0000 (21:15 +0000)]
CostModel: increase the default cost of supported floating point operations from 1 to two. Fixed a few tests that changes because now the cost of one insert + a vector operation on two doubles is lower than two scalar operations on doubles.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179413 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd debug prints.
Nadav Rotem [Fri, 12 Apr 2013 21:11:14 +0000 (21:11 +0000)]
Add debug prints.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179412 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd support for additional vector instructions in the interpreter.
Nadav Rotem [Fri, 12 Apr 2013 20:45:20 +0000 (20:45 +0000)]
Add support for additional vector instructions in the interpreter.

patch by Veselov, Yuri <Yuri.Veselov@intel.com>.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179409 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Move this logic into a static function as it's only applicable
Chad Rosier [Fri, 12 Apr 2013 20:20:54 +0000 (20:20 +0000)]
[ms-inline asm] Move this logic into a static function as it's only applicable
when parsing MS-style inline assembly.  No functional change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179407 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Address the FIXME for ImmDisp before brackets. This
Chad Rosier [Fri, 12 Apr 2013 19:51:49 +0000 (19:51 +0000)]
[ms-inline asm] Address the FIXME for ImmDisp before brackets. This
is a follow on to r179393 and r179399.  Test case to be added on
the clang side.
Part of rdar://13453209

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179403 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agolit: Fix infinite recursion when an out-of-tree test root is located inside the sourc...
Daniel Dunbar [Fri, 12 Apr 2013 19:09:09 +0000 (19:09 +0000)]
lit: Fix infinite recursion when an out-of-tree test root is located inside the source test root.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179402 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agolit: Add a test for discovery w/ test_exec_root (out-of-tree test root).
Daniel Dunbar [Fri, 12 Apr 2013 19:08:57 +0000 (19:08 +0000)]
lit: Add a test for discovery w/ test_exec_root (out-of-tree test root).

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179401 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Have the [ Symbol ] case fall into the more general logic. This
Chad Rosier [Fri, 12 Apr 2013 18:54:20 +0000 (18:54 +0000)]
[ms-inline asm] Have the [ Symbol ] case fall into the more general logic. This
is a follow on to r179393.  Test case to be added on the clang side.
Part of rdar://13453209

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179399 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoARM: Correct printing of pre-indexed operands.
Quentin Colombet [Fri, 12 Apr 2013 18:47:25 +0000 (18:47 +0000)]
ARM: Correct printing of pre-indexed operands.
According to the ARM reference manual, constant offsets are mandatory for pre-indexed addressing modes.
The MC disassembler was not obeying this when the offset is 0.
It was producing instructions like: str r0, [r1]!.
Correct syntax is: str r0, [r1, #0]!.

This change modifies the dumping of operands so that the offset is always printed, regardless of its value, when pre-indexed addressing mode is used.

Patch by Mihail Popa <Mihail.Popa@arm.com>

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179398 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Add support for operands that include both a symbol and an
Chad Rosier [Fri, 12 Apr 2013 18:21:18 +0000 (18:21 +0000)]
[ms-inline asm] Add support for operands that include both a symbol and an
immediate displacement.  Specifically, add support for generating the proper IR.
We've been able to parse this for some time now.  Test case to be added on the
clang side.
Part of rdar://13453209

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179393 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoPPC: Remove (broken) nested implicit definition lists
Hal Finkel [Fri, 12 Apr 2013 18:17:57 +0000 (18:17 +0000)]
PPC: Remove (broken) nested implicit definition lists

TableGen will not combine nested list 'let' bindings into a single list, and
instead uses only the inner scope. As a result, several instruction definitions
were missing implicit register defs that were in outer scopes. This de-nests
these scopes and makes all instructions have only one let binding which sets
implicit register definitions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179392 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd a comment about the PPC Interpretation64Bit bit
Hal Finkel [Fri, 12 Apr 2013 18:17:38 +0000 (18:17 +0000)]
Add a comment about the PPC Interpretation64Bit bit

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179391 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoHexagon: Set isPredicatedNew flag on predicate new instructions.
Jyotsna Verma [Fri, 12 Apr 2013 18:01:06 +0000 (18:01 +0000)]
Hexagon: Set isPredicatedNew flag on predicate new instructions.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179388 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoHexagon: Set isPredicatedFlase flag for all the instructions with negated predication.
Jyotsna Verma [Fri, 12 Apr 2013 17:46:52 +0000 (17:46 +0000)]
Hexagon: Set isPredicatedFlase flag for all the instructions with negated predication.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179387 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSimplify (A & ~B) in icmp if A is a power of 2
David Majnemer [Fri, 12 Apr 2013 17:25:07 +0000 (17:25 +0000)]
Simplify (A & ~B) in icmp if A is a power of 2

The transform will execute like so:
(A & ~B) == 0 --> (A & B) != 0
(A & ~B) != 0 --> (A & B) == 0

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179386 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Add the implementation for the AOK_Delete kind, which was added
Chad Rosier [Fri, 12 Apr 2013 16:26:42 +0000 (16:26 +0000)]
[ms-inline asm] Add the implementation for the AOK_Delete kind, which was added
in r179325.  Test case coming shortly on the clang side.
Part of rdar://13453209

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179383 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoLoopVectorizer: integer division is not a reduction operation
Arnold Schwaighofer [Fri, 12 Apr 2013 15:15:19 +0000 (15:15 +0000)]
LoopVectorizer: integer division is not a reduction operation

Don't classify idiv/udiv as a reduction operation. Integer division is lossy.
For example : (1 / 2) * 4 != 4/2.

Example:

int a[] = { 2, 5, 2, 2}
int x = 80;

for()
  x /= a[i];

Scalar:
  x /= 2 // = 40
  x /= 5 // = 8
  x /= 2 // = 4
  x /= 2 // = 2

Vectorized:

 <80, 1> / <2,5> //= <40,0>
 <40, 0> / <2,2> //= <20,0>

 20*0 = 0

radar://13640654

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179381 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: use full triple for ELF tests
Tim Northover [Fri, 12 Apr 2013 12:54:58 +0000 (12:54 +0000)]
AArch64: use full triple for ELF tests

These tests rely specifically on the names of ELF relocations, let alone any
other detail. There's no way they'd work if LLVM was emitting something else by
default.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179376 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAArch64: remove over-zealous use of CHECK-NEXT
Tim Northover [Fri, 12 Apr 2013 12:54:49 +0000 (12:54 +0000)]
AArch64: remove over-zealous use of CHECK-NEXT

It turns out some platforms (e.g. Windows) lay out their llvm-mc slightly
differently with extra newlines; there was no real reason for the test lines to
be consecutive, so this relaxes the FileCheck.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179375 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRevert broken pieces of r179373.
Benjamin Kramer [Fri, 12 Apr 2013 12:13:51 +0000 (12:13 +0000)]
Revert broken pieces of r179373.

You can't copy an OwningPtr, and move semantics aren't available in C++98.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179374 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoReplace uses of the deprecated std::auto_ptr with OwningPtr.
Andy Gibbs [Fri, 12 Apr 2013 10:56:28 +0000 (10:56 +0000)]
Replace uses of the deprecated std::auto_ptr with OwningPtr.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179373 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix a disconcerting bug in Value::isUsedInBasicBlock, which gave wrong answers for...
Benjamin Kramer [Fri, 12 Apr 2013 08:33:11 +0000 (08:33 +0000)]
Fix a disconcerting bug in Value::isUsedInBasicBlock, which gave wrong answers for blocks larger than 3 instrs.

Also add a unit test. PR15727.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179370 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDon't explicitly provide -pie in MSan bootstrap of LLVM, as it's now implied by the...
Alexey Samsonov [Fri, 12 Apr 2013 07:18:55 +0000 (07:18 +0000)]
Don't explicitly provide -pie in MSan bootstrap of LLVM, as it's now implied by the driver

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179367 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoTeach llvm-readobj to print ELF program headers
Nico Rieck [Fri, 12 Apr 2013 04:07:39 +0000 (04:07 +0000)]
Teach llvm-readobj to print ELF program headers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179363 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoRemove obsolete object file dumpers
Nico Rieck [Fri, 12 Apr 2013 04:07:13 +0000 (04:07 +0000)]
Remove obsolete object file dumpers

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179362 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoReplace coff-/elf-dump with llvm-readobj
Nico Rieck [Fri, 12 Apr 2013 04:06:46 +0000 (04:06 +0000)]
Replace coff-/elf-dump with llvm-readobj

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179361 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd extensive relocation tests for llvm-readobj
Nico Rieck [Fri, 12 Apr 2013 04:02:23 +0000 (04:02 +0000)]
Add extensive relocation tests for llvm-readobj

This test ensures that relocation type names returned by libObject match
the raw relocation type value.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179360 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd -expand-relocs to llvm-readobj
Nico Rieck [Fri, 12 Apr 2013 04:01:52 +0000 (04:01 +0000)]
Add -expand-relocs to llvm-readobj

This option expands shown relocations from single line to a dictionary
format:

  Relocation {
    Offset: 0x4
    Type: R_386_32 (1)
    Symbol: sym
    Info: 0x0
  }

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179359 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd missing relocation names
Nico Rieck [Fri, 12 Apr 2013 04:01:28 +0000 (04:01 +0000)]
Add missing relocation names

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179358 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoSupport MIPS64EL relocation type names
Nico Rieck [Fri, 12 Apr 2013 03:59:28 +0000 (03:59 +0000)]
Support MIPS64EL relocation type names

MIPS64EL relocation entries have up to three relocation operations. Because
libObject only exposes a single relocation name, use the concatenation of
the individual relocation type names.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179357 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd PPC instruction record forms and associated query functions
Hal Finkel [Fri, 12 Apr 2013 02:18:09 +0000 (02:18 +0000)]
Add PPC instruction record forms and associated query functions

This is prep. work for the implementation of optimizeCompare. Many PPC
instructions have 'record' forms (in almost all cases, this means that the RC
bit is set) that cause the result of the instruction to be compared with zero,
and the result of that comparison saved in a predefined condition register. In
order to add the record forms of the instructions without too much
copy-and-paste, the relevant functions have been refactored into multiclasses
which define both the record and normal forms.

Also, two TableGen-generated mapping functions have been added which allow
querying the instruction code for the record form given the normal form (and
vice versa).

No functionality change intended.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179356 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoDon't disable block layout when forcing block alignment.
Nadav Rotem [Fri, 12 Apr 2013 01:24:16 +0000 (01:24 +0000)]
Don't disable block layout when forcing block alignment.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179355 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoFix the test on linux by setting the triple and the align format
Nadav Rotem [Fri, 12 Apr 2013 01:07:16 +0000 (01:07 +0000)]
Fix the test on linux by setting the triple and the align format

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179354 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd a flag to align all basic blocks in the function.
Nadav Rotem [Fri, 12 Apr 2013 00:48:32 +0000 (00:48 +0000)]
Add a flag to align all basic blocks in the function.

When debugging performance regressions we often ask ourselves if the regression
that we see is due to poor isel/sched/ra or due to some micro-architetural
problem.  When comparing two code sequences one good way to rule out front-end
bottlenecks (and other the issues) is to force code alignment. This pass adds
a flag that forces the alignment of all of the basic blocks in the program.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179353 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAdd 179294 back, but don't use bit fields so that it works on big endian hosts.
Rafael Espindola [Fri, 12 Apr 2013 00:17:33 +0000 (00:17 +0000)]
Add 179294 back, but don't use bit fields so that it works on big endian hosts.

Original message:

Print more information about relocations.

With this patch llvm-readobj now prints if a relocation is pcrel, its length,
if it is extern and if it is scattered.

It also refactors the code a bit to use bit fields instead of shifts and
masks all over the place.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179345 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Add support for using the LENGTH, TYPE, and SIZE operators with
Chad Rosier [Thu, 11 Apr 2013 23:57:04 +0000 (23:57 +0000)]
[ms-inline asm] Add support for using the LENGTH, TYPE, and SIZE operators with
variables that use namespace alias qualifiers.  Test case coming on clang side
shortly.
Part of rdar://13499009

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179343 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Add support for using offsetof operator with variables that use
Chad Rosier [Thu, 11 Apr 2013 23:37:34 +0000 (23:37 +0000)]
[ms-inline asm] Add support for using offsetof operator with variables that use
namespace alias qualifiers.  Test case coming on clang side shortly.
Part of rdar://13499009

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179339 91177308-0d34-0410-b5e6-96231b3b80d8

11 years agoAliasing rules for struct-path aware TBAA.
Manman Ren [Thu, 11 Apr 2013 23:24:18 +0000 (23:24 +0000)]
Aliasing rules for struct-path aware TBAA.

Added PathAliases to check if two struct-path tags can alias.
Added command line option -struct-path-tbaa.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179337 91177308-0d34-0410-b5e6-96231b3b80d8

11 years ago[ms-inline asm] Pass a StringRef reference to ParseIntelVarWithQualifier so we
Chad Rosier [Thu, 11 Apr 2013 23:24:15 +0000 (23:24 +0000)]
[ms-inline asm] Pass a StringRef reference to ParseIntelVarWithQualifier so we
can build up the identifier string.  No test case as support for looking up
these type of identifiers hasn't been implemented on the clang side.
Part of rdar://13499009

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@179336 91177308-0d34-0410-b5e6-96231b3b80d8